Materials to reproduce Wallis, Funke et al. "Image content is more important than Bouma's Law for scene metamers"

# Materials to reproduce Wallis, Funke et al. "Image content is more important than Bouma's Law for scene metamers"

This repository contains the code and data used to produce the manuscript:

Wallis, T. S. A., Funke, C. M., Ecker, A. S., Gatys, L. A., Wichmann, F. A., & Bethge, M. (2018). Image content is more important than Boumas Law for scene metamers. *BioRxiv*. https://doi.org/10.1101/378521

## Experiment number labelling

The numeric labelling of experiments corresponds to chronological order in the life of this project, including pilot experiments, not to the experiments finally reported in the paper. We have kept the labelling as-is here to make the code as close to original as possible. A brief description of all experiments can be found in `experiment_overview.txt`.

- Experiment 16 is the main experiment reported in the paper (Figure 1) assessing the FS-model and CNN-model under an oddity design.
- Experiment 17 is the wedge artifact control experiment (Figure S4).
- Experiment 19 is the ABX replication (Figure S5).
- Experiment 7 is the multiscale experiment (Figure S7).
- Experiment 5 is the CNN model experiment using unique images (Figure S8).
- Experiment 6 is the CNN model experiment with repeated images (Figure S9).
- Experiment 13 is the pre-registered spatial cueing experiment (Figure S10).
- Experiment 20 is the 2IFC local texture distortion experiment (Figure S13).

## Directories and explanation

### `/code`

`/experiment/` contains matlab / PTB code and parameters specified as `yaml` files to run the experiments reported in the paper.

`/stimuli` contains code used to generate stimuli. Some of this will be specific to our cluster setup and will need re-writing to be generally useable. FS-model requires MATLAB and installation of the [model repository](http://github.com/freeman-lab/metamers) and associated toolboxes. The CNN-model is found in the ipython notebook `DNN-foveated-model.ipynb`.

`/analysis` contains code to analyse psychophysical data, including fitting Bayesian models in `brms`.

### `/figures`

Figure files used in the manuscript (as .pdf) are included here. They were assembled in Inkscape.

### `/publications`

### `/raw_data`

Contains raw psychophysical data output by each experiment. Raw eyetracking data files are not included for storage reasons and are available upon request.

### `/results`

These include the outputs of data munging (collating the separate raw data files into one file for analysis) and the saved R binaries of the fit Stan / brms models (these can be quite large).

### `/stimuli`

The stimuli used in our experiments are provided here. **Note!** The original images fall under a different license – they are part of the MIT1003 dataset. They are provided here to facilitate reproducibility.

Judd, T., Durand, F., & Torralba, A. (2012). A benchmark of computational models of saliency to predict human fixations. CSAIL Technical Reports. Retrieved from http://dspace.mit.edu/handle/1721.1/68590

Judd, T., Ehinger, K. A., Durand, F., & Torralba, A. (2009). Learning to predict where humans look. In IEEE 12th International Conference on Computer Vision (pp. 2106–2113). Kyoto. https://doi.org/10.1109/ICCV.2009.5459462

Please contact Tom Wallis (thomas punkt wallis atsymbol uni-tuebingen punkt de) if any required files are missing.

### License

Materials and code are shared here under a CC-BY license to facilitate re-use and reproducibility. **Any material belonging to a different repository remains under the license of that repository**, and is reproduced here purely for ease of use.